Summary

- 118 - where a total of 56 samples of soil moisture estimations from σ⁰HH yield RMSE and bias of 7.68 vol. % and 2.82 vol. %. Due to insignificant difference of local incidence angle between images for F32, soil moisture of higher than 100 vol. % is calculated for the field, and therefore the results for F32 are omitted as outlier.

Figure 5.19 In situ soil moisture measurement at 5–8 cm depth and Rahman estimations from σHH.

- 119 - Comparisons show overestimation of backscattering coefficients for both HH and VV polarisations in semi-empirical models universally. In particular, the Oh model in both its 2002 and 2004 versions show overestimations over 3dB for both co-polarisations, while a closer agreement is found for VV polarisation than for HH polarisation using the Dubois model. The Oh model shows large RMSE but close-to-zero bias for HV po-larisation. The semi-empirical model for ERS imagery shows the best fitted result with overestimations of less than 2dB for both co-polarisations, while no in situ surface roughness information is needed. In terms of the theoretical model the AIEM two shapes of correlation function are tested for measured s and l. It is found that the AIEM coupled with the Gaussian ACF leads to significant RMSE for backscattering coeffi-cients for both HH and VV polarisations. Similar to the Oh model, about 3dB overesti-mation of backscattering coefficients from both HH and VV polarisations are shown from the AIEM coupled with the exponential ACF. The best results are achieved by the AIEM coupled with the empirical correlation length (Baghdadi et al. 2006a), while bias is reduced to within 1dB for both co-polarisations.

After careful evaluation, all models/approaches are used for soil moisture conversion from SAR backscattering coefficients. ―Backscattering corrections‖ are adopted for those models with significant overestimating performance during the evaluation process.

Table 5.8 lists the best performance of each model/approach and their statistics in terms of RMSE and bias compared to in situ soil moisture measurements. The Dubois model is unable to provide meaningful soil moisture estimation for all samples due to its model condition limitation. An additional reason might be its limitation on fine scales (Western et al. 2001). Although the Oh model and the AIEM coupled with LUT ap-proach are able to estimate surface soil moisture with RMSE around 8 vol. %, their re-quirements of different backscattering coefficient corrections limit the applicability of these models over other test areas as an operational approach. After averaging estima-tions from both co-polarisaestima-tions, the RMSE is reduced significantly to 6.19 vol. %, equivalent to the best performance from the semi-empirical model for ERS imagery.

The semi-empirical model for ERS imagery utilises a straightforward approach and provides good results with no limited in situ measurements, i.e. only soil texture infor-mation is needed during soil moisture conversion from the dielectric constant. However, roughness factors are removed out of the model concept by consideration of average

- 120 - conditions of the test sites in the calibrated database (Loew et al. 2006a). Therefore, apart from soil moisture, there is no other perspective to assess this model‘s capability.

The AIEM coupled with the empirical correlation length is able to estimate surface soil moisture with RMSE around 6 vol. % from dual-polarisation at field scale in this study without backscattering coefficient correction. The models require in situ rms height s for model initialisation.

Clearly, large uncertainties are introduced to backscattering models through roughness parameterisation especially of the correlation length. The inaccuracy could result in scale difference between the photogrammetric technique and SAR observations. First the sufficient scale of the photogrammetric technique still needs to be reassessed to bet-ter represent the field structure. Furthermore, when the measurements of the photo-grammetric technique are upscaled to meet the SAR observation, additional errors may need to be taken into account. This in-field measurement induced inaccuracy can be overcome by the empirical correlation length. For a future operational approach, which needs roughness parameterisation, other remote sensing based roughness inversions are necessary, such as the polarimetric method (Hajnsek et al. 2003). Nevertheless, this study evaluates a number of widely used models and approaches. Results show that for a future operational approach with confident rms height available, the AIEM-Baghdadi approach is able to provide soil moisture estimation with RMSE in the order of 6 vol. %.

As an operational approach, the AIEM adapted Rahman approach has the advantage of no requirements of in situ measurements, and it can provide estimated rms height s and correlation length l for further model assessment. Note the model is calibrated on the following conditions: 0.5 cm < s < 4 cm, 1 cm < l < 40 cm, 3 vol. % < mv < 30 vol. % and 10° < θ < 40°, and therefore is only recommended be applied in these conditions.

However, the model has the following limitations. First, the approach requires at least two images with large difference of local incidence angle, i.e. larger than 20°. A failure to satisfy this requirement will lead to a failure of soil moisture estimation, as for F32 in this study. Second, since the AIEM regression requires at least three SAR images, the arbitrary selection of images may introduce errors into the final estimation. Therefore, fewer arbitrary selections are preferred. Chapter 6 will focus on solving these

limita-- 121 limita-- tions through proposing a new model for soil moisture retrieval on fields of small size in a semi-arid environment.

Table 5.8 Soil moisture conversion statistics.

Model/approach Polarisation RMSE (vol. %) Bias (vol. %)

Oh VV 8.64 2.03

HV 12.63 7.46

Semi-empirical model for ERS imagery HH 6.72 0.72

VV 6.50 0.51

HH+VV 6.21 0.11

AIEM+LUT HH 8.31 0.36

VV 8.97 1.89

HH+VV 8.15 1.67

AIEM+l_opt HH 7.30 1.27

VV 7.33 4.07

HH+VV 6.19 1.40

AIEM+Rahman HH 7.68 2.82

- 122 - There is never an end to learning. The dye

extracted from the indigo is bluer than the plant; so is the ice colder than the water.

---Xun Zi

Chapter 6 Model Development, Evaluation and